davsclaus opened a new pull request, #30018:
URL: https://github.com/apache/camel/pull/30018

   JIRA: [CAMEL-26932](https://issues.apache.org/jira/browse/CAMEL-26932), item 
7.
   
   ## The bug
   
   `camel ask` without a question opens a chat that keeps one history for all 
its questions. Every tool call and every tool result went into it in full: 
`cli_exec` alone returns up to 32 KB of stdout plus 32 KB of stderr. The whole 
history was sent with every question, and nothing trimmed it. Once it was 
larger than the context window of the model, every later question failed. A 
question that failed also stayed in the history, so the next request sent the 
failed turn again.
   
   ## The fix
   
   The chat bounds its history as the AI panel of the TUI already does:
   - a tool result is cut at 18,000 characters 
(`CatalogDocs.MAX_ANSWER_LENGTH`, so no catalog answer is cut) before it enters 
the history;
   - once a question is answered, the tool results of the questions before the 
previous one are shrunk to 400 characters, and the chat keeps the last 20 
questions. Whole turns are removed, so a tool call never loses its result;
   - a question that fails (the model is not reachable, or there is no answer 
within `--max-iterations`) is removed from the history. The note on the 
mistakes of the answer before (CAMEL-26960) is kept for the next question.
   
   The history code moved from `AiPanel` (TUI) to camel-jbang-core as 
`ChatHistory`, which `camel ask` and the AI panel now share. `AiPanel`'s 
methods call it, so its tests are unchanged. Its size limit in `SourceSizeTest` 
goes down to 3093 lines.
   
   ## Tests
   
   - `AskLoopTest.aQuestionOfAChatThatFailsLeavesNoTraceInTheHistory`: a 500 
from the model leaves the history empty, and the next request has only the 
system prompt and the new question.
   - `AskLoopTest.theToolResultsOfAChatStayBounded`: three questions with a 
36,000-character tool result each. All results are cut, the result of the first 
question is shrunk, and the previous question's result is kept.
   - Both fail on main. The core suite and the TUI suite (2617 tests) are 
green. `DependencyUpdateTest` failed once and passed on rerun; it fails once 
per run on every branch.
   
   Docs: `camel-jbang-ai.adoc` says how the chat keeps its history bounded.
   
   _Claude Code on behalf of davsclaus_
   
   🤖 Generated with [Claude Code](https://claude.com/claude-code)
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to